Paper deep dive

A Context-Aware Cultural Heritage Guide Powered by LLMs

Liliana Ardissono, Fabio Ferrero, Angelo Geninatti Cossatin, Claudio Mattutino, Noemi Mauro

Year: 2026Venue: arXiv preprintArea: cs.IRType: PreprintEmbeddings: 19

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 8/4/2026, 4:37:53 AM

Summary

The paper presents an extension of the 'Triangolazioni' Cultural Heritage (CH) web application that integrates Large Language Models (LLMs) to provide context-aware, personalized information. The system uses a loosely-coupled architecture where a chatbot, powered by the Perplexity Sonar model (based on Llama 3.3), interacts with users via a web interface. It tracks user navigation to maintain a shared context, enabling the LLM to generate personalized question suggestions and answers based on the user's browsing history and the curated content of the CH site. This approach overcomes limitations of traditional chatbots by decoupling the LLM from the knowledge base and supporting flexible LLM replacement.

Entities (11)

Large Language Models · technology · 99%Egyptian Museum · point-of-interest · 98%Measurements and Revolutions · thematic-path · 98%Triangolazioni · software-system · 98%Turin · location · 98%Llama-3.3-70B · software-model · 95%Omeka-S · software-system · 95%Perplexity Sonar · software-model · 95%Context (CTX) · data-structure · 90%Simulated Chatbot (SC) · software-component · 90%Simulated User (SU) · software-component · 90%

Relation Signals (9)

Perplexity Sonar → builton → Llama-3.3-70B

confidence 95% · which is built on top of Llama 3.3 70B.

Triangolazioni → extends → LLM Integration

confidence 95% · We present an extension of Triangolazioni (a Cultural Heritage webapp) to enrich curated content with context-dependent, external information provided by Large Language Models (LLMs)

LLM → generates → Context-Dependent Answers

confidence 95% · The system answers questions in the widget “Write your own question”... answers the inquiries context-dependently.

LLM → generates → Personalized Questions

confidence 95% · The chatbot also proposes 3 personalized generated questions to suggest content exploration in greater depth.

Perplexity Sonar → powers → Chatbot

confidence 95% · The chatbot is powered by the sonar model provided by Perplexity

Omeka-S → stores → Curated Data

confidence 95% · employs the Omeka-S semantic content management system to store and manage curated data.

Simulated User (SU) → mediates → Interaction

confidence 90% · SU and SC operate within the web browser and mediate the exchange of information between the website and the chatbot.

Simulated Chatbot (SC) → mediates →

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:We present an extension of Triangolazioni (a Cultural Heritage webapp) to enrich curated content with context-dependent, external information provided by Large Language Models (LLMs) within a loosely-coupled architecture agnostic to the LLM. The system supports context-dependent information search and presentation within an architecture agnostic to the exploited LLM.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

17,412 characters extracted from source content.

Expand or collapse full text

A Context-Aware Cultural Heritage Guide Powered by LLMs Liliana Ardissono, Fabio Ferrero, Angelo Geninatti Cossatin, Claudio Mattutino, Noemi Mauro This is the authors’ version of the work. The definitive Version of Record was published in Proceedings of the 34th ACM Conference on User Modeling, Adaptation and Personalization (UMAP ’26), June 08–11, 2026, Gothenburg, Sweden. DOI: https://doi.org/10.1145/3774935.3812714 © Licensed under C BY-NC-ND 4.0. A Context-Aware Cultural Heritage Guide Powered by LLMs Liliana Ardissono 0000-0002-1339-4243 Università degli Studi di TorinoCorso Svizzera 185TorinoItaly10149 liliana.ardissono@unito.it , Fabio Ferrero 0009-0007-2941-5030 Università degli Studi di TorinoCorso Svizzera 185TorinoItaly10149 fab.ferrero@unito.it , Angelo Geninatti Cossatin 0009-0007-5378-7061 Università degli Studi di TorinoCorso Svizzera 185TorinoItaly10149 angelo.geninatticossatin@unito.it , Claudio Mattutino 0000-0002-0413-2436 Università degli Studi di TorinoCorso Svizzera 185TorinoItaly10149 claudio.mattutino@unito.it and Noemi Mauro 0000-0001-8234-3266 Università degli Studi di TorinoCorso Svizzera 185TorinoItaly10149 noemi.mauro@unito.it Abstract. We present an extension of Triangolazioni (a Cultural Heritage webapp) to enrich curated content with context-dependent, external information provided by Large Language Models (LLMs) within a loosely-coupled architecture agnostic to the LLM. The system supports context-dependent information search and presentation within an architecture agnostic to the exploited LLM. Context-aware information provision in CH websites, LLMs †copyright: none†ccs: Information systems Web searching and information discovery†ccs: Human-centered computing Interaction techniques†footnotetext: © 2026 Copyright held by the owner/author(s). Licensed under C BY-NC-ND 4.0. Published at UMAP ’26: Proceedings of the 34th ACM Conference on User Modeling, Adaptation and Personalization, DOI: https://doi.org/10.1145/3774935.3812714 1. Introduction Large Language Models (LLMs) open new possibilities for developing Cultural Heritage (CH) guides based on chatbots, thanks to their strength in understanding and answering users’ questions, and their ability to gather information from heterogeneous data sources. However, they limit interaction with users to message-based conversations. Moreover, they have to be frequently replaced to keep pace with the quick evolution of Generative AI technology towards more reliable models. Early research on Cultural Heritage exploration investigated advanced information filtering and presentation techniques to enhance user experience when navigating curated physical or virtual CH sites (Kuflik et al., 2011; De Carolis et al., 2018). Chatbots have been introduced in CH sites to simplify the interaction with the user, but they work on specific knowledge bases. For example, (Machidon et al., 2020) proposed a chatbot based on Google’s DialogFlow to assist users in exploring the content of the Europeana ontology. To expand the available information, external data sources have been integrated through Semantic Web technologies (Faralli et al., 2022; Kim et al., 2017; Rinaldi et al., 2022). However, this integration still limits the knowledge bases to a closed set of resources. LLMs can acquire information from public repositories to extend their knowledge. However, they are rarely used in CH systems. Moreover, systems that leverage this technology (e.g., (Trichopoulos et al., 2023)) raise two main concerns. First, as the LLM represents the user interface, it constrains the interaction with the user to the exchange of messages. Second, the LLM is typically fine-tuned to the system’s knowledge base to enhance its factual accuracy. Thus, the system is tied to it. In contrast, our work integrates web-based presentation and conversation with the LLM within a loosely-coupled architecture to facilitate LLM replacement. This paper presents a system that integrates a web-based cultural heritage site with an LLM-powered chatbot that extends curated content with external data, enriching the information presented beyond closed knowledge bases. By tracking the user’s navigation of the website, the system supports content exploration through question formulation support and personalized question-answering. The system architecture, developed for (Geninatti Cossatin et al., 2025), provides a loosely-coupled integration of LLMs into a CH website. It synchronizes the chatbot with the user’s browsing activity in the website through a shared context that makes the system agnostic to the LLM. The novel features of our work are: (1) A context-aware question-answering function based on LLMs to satisfy individual information needs beyond a CH website’s curated content. (2) A context-dependent suggestion of questions to help the user expand the visited content with relevant information. (3) A loosely coupled architectural model to flexibly integrate LLMs in a CH website. A video of the interaction with the system is available at the following link: https://youtu.be/Vf_5yhndMGk. (a) Presentation of a thematic path. Measurements and Revolutions thematic path The upper portion of the page shows the thematic path on a geographical map, with the relevant Points of Interest, and describes the thematic path as a story. Below, a graphical component includes a link to each of these Points of Interest, with a small image, a short text, and a button to open the details page. (b) Suggested questions and chatbox’s answer to the selected one. Figure 1. The user interface within the path Measurements and Revolutions (visualization for medium size screens). Measurements and Revolutions - Egyptian Museum The upper portion of the page shows an image of the entrance of the museum. Below, there is a description of the museum and the graphical component showing the chatbot. This component includes three suggested questions and a text area to chat with the user. 2. The system We build on the Triangolazioni CH guide (Mauro et al., 2022), which presents thematic paths describing Points of Interest in Turin and its surroundings. Triangolazioni organizes thematic path as geolocalized narratives. It describes their places, historical personages, monuments, and historical or artistic objects. Its user interface is responsive and can be accessed from desktop and mobile devices. Our system extends Triangolazioni by integrating a chatbot powered by an LLM into its web pages.†The knowledge base of the original Triangolazioni guide is in Italian, and for the demonstration, we translated a part of it into English. Figure 1 shows two sample pages. The left one describes the “Measurements and Revolutions” thematic path, which focuses on measurement units (for time, space, etc.) and their evolution across time. The right one presents a Point of Interest (“Egyptian Museum”) in the context of the thematic path. The chatbot works in the “Keep exploring:” area of the web page. An input field enables sending free-text questions (“Please write here your question to delve into this topic…”). The chatbot also proposes 3 personalized generated questions to suggest content exploration in greater depth. The user can write a question or submit the sample questions by clicking the “Ask” button, as in Figure 1(b). The system answers questions in the widget “Write your own question”. Below each answer, it shows another input field to let the user submit further questions, enabling a longer dialog. The chatbot knows about the content of the current page and the previously browsed ones and answers the inquiries context-dependently. Therefore, both the suggested questions and the generated responses are personalized to the user’s browsing history in the mobile guide. Our system tracks the information that the user receives when interacting with it. By sharing this context with the chatbot, it instructs the underlying LLM to generate personalized questions and answers. For example, Figure 1(b) shows the questions suggested in an interaction where the user only explored the “Measurements and Revolutions” path: “How did ancient Egyptians calibrate their royal cubits for precise construction?”, “What revolutionary role did the henu play in Egyptian volume measurements?”, and “How might Turin’s unguentarium reveal shifts from ancient to modern metrics?”. In a different interaction, the user first explored some parts of the thematic path “Measuring the Earth” focusing on mathematical aspects of spatial measurement. In that case, when moving to “Measurements and Revolutions” and inspecting the “Egyptian Museum”, the suggested questions were: “What ancient Egyptian tools in the museum helped standardize grain measurements?”, “How did the royal cubit influence revolutionary changes in metric systems?”, “Which papyrus artifact reveals Egypt’s role in early revolutionary metrology?”. These questions have a broader scope, referring to the standardization of measures over time. 3. Architecture Our system is implemented using HTML and JavaScript and employs the Omeka-S semantic content management system†https://omeka.org/s/) to store and manage curated data. The chatbot is powered by the sonar model provided by Perplexity†https://w.perplexity.ai/api-platform/resources/meet-new-sonar , which is built on top of Llama 3.3 70B. We used a temperature of 0 to ensure reproducibility. It is integrated through implicit prompts shown in Table 1. Table 1. Prompt templates to interact with the chatbot. Prompt templates to interact with the chatbot There are five prompt templates. In each template, the variable part is enclosed in brackets. P1 Hello, I am your digital tour guide for Turin. How can I assist you? P2a I am beginning this thematic path in the Turin area: [NAME OF THE PATH]. Please tell me something about it. P2b I am looking at this place in the Turin area: [NAME OF THE PLACE] in the path [NAME OF THE PATH]. Please tell me something about it. P3 You are a tour guide. Please write 3 very short and engaging questions about [PLACE NAME] in the thematic path [PATH NAME], based on our conversation so far, but whose answer is not contained in it. Please, each line should contain only the question, without quotation marks. P4 You are a tour guide of Turin. Please briefly answer the following question, without repeating what we have said in this conversation: [QUESTION] To align the chatbot’s behavior with the user’s navigation of the website, the system maintains an interaction context that captures (i) the information presented to the user during web browsing and (i) the responses previously generated by the chatbot. We represent this interaction context (hereafter denoted as CTX) as a conversation between the user and the system: the user’s browsing activity is modeled as a dialog where the user asks for information and the guide responds by presenting the requested web pages. These pages include curated, static content and the dynamic content generated by the chatbot. CTX includes dialog turns representing the user’s questions for the chatbot (LLM) and its responses. Since the chatbot itself is stateless, an external software component is responsible for constructing and maintaining CTX. This component tracks both browsing actions and question-answer interactions. It supplies the full context to the chatbot at each invocation. The maximum size of CTX is constrained by the context window of the underlying LLM. The integration between the web-based guide and the chatbot follows a loosely coupled architecture. A simulated user (SU) and a simulated chatbot (SC) act as proxies for the human user and the LLM-powered chatbot, respectively. SU and SC operate within the web browser and mediate the exchange of information between the website and the chatbot. They merge the content of visited web pages with the chatbot’s generated outputs into the shared conversation context CTX. The interaction flow proceeds as follows: (1) When the user accesses the website for the first time, CTX is initialized with a welcome dialog act generated by the simulated chatbot SC and addressed to the simulated user SU (prompt P1 in Table 1). (2) Each time the user navigates to a web page x: (a) The user interface provides SU with the URL and title of x, which identifies a topic T (e.g., “Egyptian Museum”). (b) To model the user’s intent, SU invokes SC using a prompt expressing interest in learning more about T. Prompt P2a/2b is used for pages describing thematic paths/Points of Interest. (c) SC retrieves the content of page x from its URL and appends it to CTX. This step unifies web navigation and question–answering by treating the page content as if it were provided by the LLM in response to SU’s request. As a result, CTX accurately reflects the information received by the user. (d) To enable the generation of contextual question suggestions, SU asks SC to produce three candidate questions based on the current CTX (prompt P3). (e) SC forwards the prompt together with CTX to the LLM through its API. In turn, SC collects the generated questions and returns them to SU. Finally, SU forwards the questions to the user interface, where they are displayed to the user. For brevity, we omit the description of interactions triggered by explicit user questions and refer the reader to (Geninatti Cossatin et al., 2025) for further details. 4. Conclusions In a user study (Geninatti Cossatin et al., 2025), we found that our system has higher usability than the original Triangolazioni app. Moreover, its context-dependent question suggestions enhance user experience compared to free-text interaction with the LLM, especially for users with low levels in the Curiosity and Exploration Inventory-I (Kashdan et al., 2009). This research extends the work of the Triangolazioni Project (w.triangolazioni.unito.it/), and has been funded by our University (Grant for Internationalization). Code and data are available at https://anonymous.4open.science/r/tell-me-more-A4E5/. References (1) De Carolis et al. (2018) Berardina De Carolis, Cristina Gena, Tsvi Kuflik, and Joel Lanir. 2018. Special issue on advanced interfaces for cultural heritage. International Journal of Human-Computer Studies 114 (2018), 1–2. https://doi.org/10.1016/j.ijhcs.2018.02.007 Faralli et al. (2022) Stefano Faralli, Andrea Lenzi, and Paola Velardi. 2022. A Large Interlinked Knowledge Graph of the Italian Cultural Heritage. In Proceedings of the Thirteenth Language Resources and Evaluation Conference. European Language Resources Association, Marseille, France, 6280–6289. https://aclanthology.org/2022.lrec-1.675 Geninatti Cossatin et al. (2025) Angelo Geninatti Cossatin, Noemi Mauro, Fabio Ferrero, and Liliana Ardissono. 2025. Tell me more: integrating LLMs in a cultural heritage website for advanced information exploration support. Information Technology & Tourism (2025). https://doi.org/10.1007/s40558-025-00312-8 Kashdan et al. (2009) Todd Kashdan, Matthew Gallagher, Paul Silvia, Beate Winterstein, William Breen, Daniel Terhar, and Michael Steger. 2009. The Curiosity and Exploration Inventory-I: Development, Factor Structure, and Psychometrics. Journal of research in personality 43 (12 2009), 987–998. https://doi.org/10.1016/j.jrp.2009.04.011 Kim et al. (2017) Hayun Kim, Tamás Matuszka, Jea-In Kim, Jungwha Kim, and Woontack Woo. 2017. Ontology-Based Mobile Augmented Reality in Cultural Heritage Sites: Information Modeling and User Study. Multimedia Tools and Applplications 76, 24 (dec 2017), 26001–26029. https://doi.org/10.1007/s11042-017-4868-6 Kuflik et al. (2011) Tsvi Kuflik, Oliviero Stock, Massimo Zancanaro, Ariel Gorfinkel, Sadek Jbara, Shahar Kats, Julia Sheidin, and Nadav Kashtan. 2011. A Visitor’s Guide in an “Active Museum”: Presentations, Communications, and Reflection. ACM Journal of Computing and Cultural Heritage 3, 3 (2011), 175–209. https://doi.org/10.1145/1921614.1921618 Machidon et al. (2020) Octavian-Mihai Machidon, Aleš Tavčar, Matjaž Gams, and Mihai Duguleană. 2020. CulturalERICA: A conversational agent improving the exploration of European cultural heritage. Journal of Cultural Heritage 41 (2020), 152–165. https://doi.org/10.1016/j.culher.2019.07.010 Mauro et al. (2022) Noemi Mauro, Angelo Geninatti Cossatin, Ester Cravero, Liliana Ardissono, Guido Magnano, Marco Giardino, and Claudio Mattutino. 2022. A Mobile Guide to Explore Interconnections between Science, Art and Territory. In Adjunct Proceedings of the 30th ACM Conference on User Modeling, Adaptation and Personalization (UMAP ’22 Adjunct). ACM, Barcelona, Spain, 397–401. https://doi.org/10.1145/3511047.3537649 Rinaldi et al. (2022) Antonio M. Rinaldi, Cristiano Russo, and Cristian Tommasino. 2022. An Approach Based on Linked Open Data and Augmented Reality for Cultural Heritage Content-Based Information Retrieval. In Computational Science and Its Applications – ICCSA 2022, Osvaldo Gervasi, Beniamino Murgante, Eligius M. T. Hendrix, David Taniar, and Bernady O. Apduhan (Eds.). Springer International Publishing, Cham, 99–112. https://doi.org/10.1007/978-3-031-10450-3_8 Trichopoulos et al. (2023) Georgios Trichopoulos, Markos Konstantakis, George Caridakis, Akrivi Katifori, and Myrto Koukouli. 2023. Crafting a Museum Guide Using ChatGPT4. Big Data and Cognitive Computing 7, 3 (2023). https://doi.org/10.3390/bdcc7030148

Instant research discovery